A robust, containerised workflow for processing multi-session, ultra high-field, high-res MRI data

2264 

Abstract Submission 

Sriranga Kashyap¹, Carlos Gomes², Nikolai Axmacher², Kamil Uludag^1,3,4

¹Krembil Brain Institute, University Health Network, Toronto, Ontario, Canada, ²Department of Neuropsychology, Ruhr University Bochum, Bochum, North Rhine-Westphalia, Germany, ³Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada, ⁴Center for Neuroscience Imaging Research, Institute for Basic Science & Department of Biomedical Engineering, Sungkyunkwan University, Suwon, Korea, Republic of

Sriranga Kashyap, Ph.D.  
Krembil Brain Institute, University Health Network
Toronto, Ontario, Canada

Carlos Gomes, Ph.D.  
Department of Neuropsychology, Ruhr University Bochum
Bochum, North Rhine-Westphalia, Germany

Nikolai Axmacher  
Department of Neuropsychology, Ruhr University Bochum
Bochum, North Rhine-Westphalia, Germany

Kamil Uludag  
Krembil Brain Institute, University Health Network|Department of Medical Biophysics, University of Toronto|Center for Neuroscience Imaging Research, Institute for Basic Science & Department of Biomedical Engineering, Sungkyunkwan University
Toronto, Ontario, Canada|Toronto, Ontario, Canada|Suwon, Korea, Republic of

With 7T MRI, we are able to image the human brain anatomy and function at a submillimetre scale. However, high spatial resolution (high-res) fMRI is often limited to partial brain coverage. For example, the majority of layer-fMRI studies have focused on primary cortices due to well-established correspondence between cortical depths and the neuronal microcircuit [1]. Recently, interest has grown to study non-primary brain areas and higher-level cognitive processes with layer-fMRI, e.g., using language processing, episodic and working memory tasks [2,3,4]. An important challenge for this endeavour is the fact that there is a noticeable lack of readily-available and generalisable pipelines that can effectively preprocess 7T high-res fMRI data, particularly when small fields-of-view are involved and data are acquired over multiple sessions. Here, we present a workflow that has been developed for preprocessing of small field-of-view, multi-session, sub-millimetre resolution fMRI data of non-neocortical structures, such as the hippocampus. This workflow, building upon our previous work [5], is modular, containerised and BIDS-compliant, requiring only minimal input from the user. The workflow primarily uses ANTs [6] for a majority of registration steps and integrates standard software, such as Freesurfer, SPM, AFNI and FSL for various intermediate steps.

We used four high-res fMRI datasets from an ongoing spatial navigation study acquired on Siemens Terra 7T at the ELH Institute, Essen, Germany. Data were acquired in 2 sessions: in ses-01, 0.75 mm iso MP2RAGE, hippocampus-aligned 2D-TSE (0.44 x 0.44 x 1.5 mm3, 3 reps), and 2 fMRI runs (func, 0.8 mm iso 3D-EPI) and 4 fMRI runs were acquired in ses-02 (6 fMRI runs per subject). Opposite phase-encoded (oPE) data were acquired after each fMRI run. Subjects performed a spatial navigation task [7]. Briefly, our workflow (Fig1) realigns 'func' and 'oPE' data and performs distortion correction using ANTs (like 3dQwarp) for each run in each session. Inter-session alignment is done using the corrected runwise data and 'func2anat' is estimated using the boundary-based registration (BBR) algorithm [8] in FSL, as it outperforms typical cost functions in data with the signal dropouts and artefacts (e.g., in ventral/medial temporal lobes). All matrices and warps are preserved and are concatenated and applied in a single resampling step to reduce interpolation errors. Motion and QC plots are generated using Python.

For all subjects, both the structural similarity index (intra-session: 0.92 ± 0.023, inter-session: 0.873 ± 0.019; 1.0 being identical) and the normalised root mean squared error (intra-session: 0.041 ± 0.008, inter-session: 0.065 ± 0.027; 0 being identical) indicate a high degree of similarity between the reference and resampled images (both mean EPI, Fig. 2) after processing through our workflow.

Inter-session registration is more challenging than intra-session as different head positions and orientations have different distortions, and regions of dropouts and artefacts, and high-res fMRI is more sensitive to these. Improving on current state-of-the-art workflows, our approach combines optimal data acquisition (e.g., runwise oPE with fMRI) with automated, high-res optimised image processing (e.g., cost-function choice, ROI-confined registrations, image filtering to enhance boundaries) and one-step resampling, to ensure accurate intra- and inter-session results. Taking advantage of efficient, multi-threaded tools and evaluating on a system with just 32 Gb of RAM (typical University PC), the processing times were kept reasonable around 32 minutes per fMRI run (HPCs and servers will be faster) for high-res fMRI with partial brain coverage and large file sizes (900 MB, compressed). Containerisation using Docker takes care of software dependencies and helps easy adoption of our integrative workflow by the wider neuroscience community.

Motion Correction and Preprocessing ²

Workflows ¹

Informatics Other

Cortical Layers

Data analysis

FUNCTIONAL MRI

HIGH FIELD MR

MRI

Workflows